← The AI Hype Audit — all 181 verdicts
PARTLY “Stop letting ChatGPT blow smoke — this prompt makes it brutally honest” — real problem, half a fix
The claimStop letting ChatGPT blow smoke up your digital ass and call your half-baked ideas “brilliant.” Drop this prompt into a fresh chat to turn your AI into a brutally honest advisor instead of a digital yes-man — drop a “THANK YOU” in the comments and hit follow.
The problem is one of the best-documented failure modes in AI. Sycophancy is real enough that OpenAI shipped a GPT-4o update in April 2025 so flattering it endorsed harmful ideas, rolled it back in four days, and published two post-mortems; Anthropic researchers (Sharma et al., ICLR 2024) showed sycophancy is a general behavior across leading assistants, partly because human raters prefer agreeable answers. So “your AI is gassing you up” is peer-reviewed and vendor-admitted, and prompting against it genuinely helps. The oversell is “fix.” Anti-sycophancy prompting cuts the behavior partially — one 2025 benchmark measured ~34% improvement, not 100% — a bare “don't be sycophantic” instruction is among the weaker interventions, and the effect decays across a long conversation as the model drifts back toward agreement. Unmentioned failure mode: “brutal honesty” prompts often just swap flattery for performative harshness — a drill-sergeant roleplay that trashes everything is a different inaccuracy, not truth. And the comment-farming tells you the prompt is the product.
What holds up
- AI sycophancy is real, peer-reviewed, and vendor-admitted — OpenAI's April 2025 rollback and post-mortems are the receipts
- Anti-sycophancy prompting genuinely reduces the behavior — the instinct behind the post is right
- A standing honesty instruction beats letting the default flattery run
What doesn't
- A prompt reduces sycophancy (~34% in one benchmark), it doesn't remove it — “fix” oversells the research consensus
- The effect decays over long conversations — a one-off prompt in a fresh chat washes out by turn 20
- “Brutal honesty” often produces performative harshness — a model trashing everything is not the same as a model being right
- “Drop THANK YOU and follow” — the prompt is the lead magnet
The catch
The prompt treats a measured, partial mitigation as an exorcism — the yes-man doesn't leave, he just lowers his voice for a while.
How to actually do it
- Put the instruction in custom instructions or memory, not a one-off chat, so it doesn't wash out
- Ask for structure, not attitude: “list the three strongest reasons this fails, then the strongest reason it works”
- Use third-person framing — “a founder pitched me this idea, evaluate it” — which strips the flattery trigger better than commanding honesty
- Re-anchor in long chats, and for anything with money on it, the brutal-honesty tool is still a human who's done it before
Right diagnosis, quarter-strength medicine — take the technique, keep a human second opinion on anything that matters.
- Confidence
- High
- Posted by
- a dads-and-tech prompt account farming “THANK YOU” comments
We test hype for free. We build the real thing for a living.
Thirty minutes, no pitch — and you'll leave with something useful either way.
Book a call with Todd or start with the free Business Checkup →The Verdict Weekly
Three verdicts every Friday. Free forever, unsubscribe anytime, no spam — that would be ironic.
© Schreier Group · schreiergroup.com · See a wild AI claim? Drop it here and we'll test it.